KernelGenBench-MS (AutoKernel): leaderboard
KernelGenBench asks models and coding agents to write Triton kernels that drop into PyTorch in place of an existing operator. A kernel counts only when it passes numerical checks across a combinatorial suite of shapes, data types and layouts and clears anti-hack interception that blocks calls back to the reference implementation. This board covers all 210 multi-source operators on an NVIDIA A100, generated by the kernel-specialized AutoKernel agent with execution feedback.
Metric: Clean pass rate (%; share of the 210 KernelGenBench-MS operators (110 PyTorch ATen, 50 vLLM, 50 cuBLAS) whose generated Triton kernel passes numerical validation on every test case of its combinatorial shape, dtype and layout suite and every anti-hack check, as a drop-in replacement; the kernel-specialized AutoKernel agent with execution feedback; NVIDIA A100; 30-minute budget per operator). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | AutoKernel + GLM-5 | 71 | |
| 2 | AutoKernel + Qwen3.5-27B | 47 | |
| 3 | AutoKernel + MiniMax-M2.5 | 43 |
Interactive version: theaggregate.ai/benchmark?slug=kernelgenbench-ms-autokernel · How It Works · Data refreshed daily, snapshot 2026-11-13.